> The gulf is bridgeable.

Only with an LLM that's actually at agent-quality.

If "useful chatbot" and "useful agent" are two rungs on a ladder, the rung before them is "useful autocomplete". Autocomplete that only gets the next token right 90% of the time won't give you compiling code.

How many of these layers are just trying to rediscover/rebuild the idempotence of code?